This research explores whether language models have acquired knowledge about physical vowel articulation without explicit training on the subject.

• LMs demonstrate surprising understanding of tongue positions for different vowels
• Models can accurately map relationships between vowel sounds and articulation mechanics
• Performance varies across languages and model sizes, with larger models showing better comprehension
• Results show potential applications for pronunciation training and linguistic education

This research matters for linguistics by revealing how artificial intelligence might develop implicit understanding of physical speech production—bridging computational linguistics with traditional phonetics.

Tonguescape: Exploring Language Models Understanding of Vowel Articulation